[DAGCombiner] Use the right thumbv7meb triple for ARM big-endian test.

1 files changed, 71 insertions, 93 deletions

diff --git a/llvm/test/CodeGen/ARM/load-combine-big-endian.ll b/llvm/test/CodeGen/ARM/load-combine-big-endian.ll
index 7b028714f4f..d045f1f96ee 100644
--- a/llvm/test/CodeGen/ARM/load-combine-big-endian.ll
+++ b/llvm/test/CodeGen/ARM/load-combine-big-endian.ll
@@ -1,8 +1,8 @@
 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
 ; RUN: llc < %s -mtriple=armeb-unknown | FileCheck %s
 ; RUN: llc < %s -mtriple=armv6eb-unknown | FileCheck %s --check-prefix=CHECK-ARMv6
-; RUN: llc < %s -mtriple=thumbv6m-none-eabi | FileCheck %s --check-prefix=CHECK-THUMBv6
-; RUN: llc < %s -mtriple=thumbv7m-none-eabi | FileCheck %s --check-prefix=CHECK-THUMBv7
+; RUN: llc < %s -mtriple=thumbv6meb-none-eabi | FileCheck %s --check-prefix=CHECK-THUMBv6
+; RUN: llc < %s -mtriple=thumbv6meb-none-eabi | FileCheck %s --check-prefix=CHECK-THUMBv7
 
 ; i8* p; // p is 4 byte aligned
 ; ((i32) p[0] << 24) | ((i32) p[1] << 16) | ((i32) p[2] << 8) | (i32) p[3]
@@ -20,13 +20,11 @@ define i32 @load_i32_by_i8_big_endian(i32* %arg) {
 ; CHECK-THUMBv6-LABEL: load_i32_by_i8_big_endian:
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
-; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_big_endian:
 ; CHECK-THUMBv7:       @ %bb.0:
 ; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
-; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -75,11 +73,13 @@ define i32 @load_i32_by_i8_bswap(i32* %arg) {
 ; CHECK-THUMBv6-LABEL: load_i32_by_i8_bswap:
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_bswap:
 ; CHECK-THUMBv7:       @ %bb.0:
 ; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
+; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -120,13 +120,11 @@ define i32 @load_i32_by_i16_by_i8_big_endian(i32* %arg) {
 ; CHECK-THUMBv6-LABEL: load_i32_by_i16_by_i8_big_endian:
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
-; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i16_by_i8_big_endian:
 ; CHECK-THUMBv7:       @ %bb.0:
 ; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
-; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -167,20 +165,12 @@ define i32 @load_i32_by_i16(i32* %arg) {
 ;
 ; CHECK-THUMBv6-LABEL: load_i32_by_i16:
 ; CHECK-THUMBv6:       @ %bb.0:
-; CHECK-THUMBv6-NEXT:    ldrb r1, [r0, #2]
-; CHECK-THUMBv6-NEXT:    ldrb r2, [r0, #3]
-; CHECK-THUMBv6-NEXT:    lsls r2, r2, #8
-; CHECK-THUMBv6-NEXT:    adds r1, r2, r1
-; CHECK-THUMBv6-NEXT:    ldrh r0, [r0]
-; CHECK-THUMBv6-NEXT:    lsls r0, r0, #16
-; CHECK-THUMBv6-NEXT:    adds r0, r0, r1
+; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i16:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrh r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrh r0, [r0, #2]
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #16
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i16*
@@ -210,22 +200,12 @@ define i32 @load_i32_by_i16_i8(i32* %arg) {
 ;
 ; CHECK-THUMBv6-LABEL: load_i32_by_i16_i8:
 ; CHECK-THUMBv6:       @ %bb.0:
-; CHECK-THUMBv6-NEXT:    ldrb r1, [r0, #3]
-; CHECK-THUMBv6-NEXT:    ldrb r2, [r0, #2]
-; CHECK-THUMBv6-NEXT:    lsls r2, r2, #8
-; CHECK-THUMBv6-NEXT:    adds r1, r2, r1
-; CHECK-THUMBv6-NEXT:    ldrh r0, [r0]
-; CHECK-THUMBv6-NEXT:    lsls r0, r0, #16
-; CHECK-THUMBv6-NEXT:    adds r0, r1, r0
+; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i16_i8:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrb r2, [r0, #2]
-; CHECK-THUMBv7-NEXT:    ldrh r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #3]
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r2, lsl #8
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #16
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i16*
@@ -277,14 +257,18 @@ define i64 @load_i64_by_i8_bswap(i64* %arg) {
 ;
 ; CHECK-THUMBv6-LABEL: load_i64_by_i8_bswap:
 ; CHECK-THUMBv6:       @ %bb.0:
-; CHECK-THUMBv6-NEXT:    ldr r2, [r0]
-; CHECK-THUMBv6-NEXT:    ldr r1, [r0, #4]
-; CHECK-THUMBv6-NEXT:    mov r0, r2
+; CHECK-THUMBv6-NEXT:    ldr r1, [r0]
+; CHECK-THUMBv6-NEXT:    ldr r0, [r0, #4]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
+; CHECK-THUMBv6-NEXT:    rev r1, r1
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i64_by_i8_bswap:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrd r0, r1, [r0]
+; CHECK-THUMBv7-NEXT:    ldr r1, [r0]
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0, #4]
+; CHECK-THUMBv7-NEXT:    rev r0, r0
+; CHECK-THUMBv7-NEXT:    rev r1, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i64* %arg to i8*
@@ -345,17 +329,16 @@ define i64 @load_i64_by_i8(i64* %arg) {
 ;
 ; CHECK-THUMBv6-LABEL: load_i64_by_i8:
 ; CHECK-THUMBv6:       @ %bb.0:
-; CHECK-THUMBv6-NEXT:    ldr r1, [r0]
-; CHECK-THUMBv6-NEXT:    ldr r0, [r0, #4]
-; CHECK-THUMBv6-NEXT:    rev r0, r0
-; CHECK-THUMBv6-NEXT:    rev r1, r1
+; CHECK-THUMBv6-NEXT:    ldr r2, [r0]
+; CHECK-THUMBv6-NEXT:    ldr r1, [r0, #4]
+; CHECK-THUMBv6-NEXT:    mov r0, r2
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i64_by_i8:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrd r1, r0, [r0]
-; CHECK-THUMBv7-NEXT:    rev r0, r0
-; CHECK-THUMBv7-NEXT:    rev r1, r1
+; CHECK-THUMBv7-NEXT:    ldr r2, [r0]
+; CHECK-THUMBv7-NEXT:    ldr r1, [r0, #4]
+; CHECK-THUMBv7-NEXT:    mov r0, r2
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i64* %arg to i8*
@@ -424,11 +407,14 @@ define i32 @load_i32_by_i8_nonzero_offset(i32* %arg) {
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    movs r1, #1
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0, r1]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_nonzero_offset:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldr.w r0, [r0, #1]
+; CHECK-THUMBv7-NEXT:    movs r1, #1
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0, r1]
+; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
 
 
@@ -479,11 +465,14 @@ define i32 @load_i32_by_i8_neg_offset(i32* %arg) {
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    subs r0, r0, #4
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_neg_offset:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldr r0, [r0, #-4]
+; CHECK-THUMBv7-NEXT:    subs r0, r0, #4
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
+; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
 
 
@@ -526,13 +515,12 @@ define i32 @load_i32_by_i8_nonzero_offset_bswap(i32* %arg) {
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    movs r1, #1
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0, r1]
-; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_nonzero_offset_bswap:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldr.w r0, [r0, #1]
-; CHECK-THUMBv7-NEXT:    rev r0, r0
+; CHECK-THUMBv7-NEXT:    movs r1, #1
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0, r1]
 ; CHECK-THUMBv7-NEXT:    bx lr
 
 
@@ -575,13 +563,12 @@ define i32 @load_i32_by_i8_neg_offset_bswap(i32* %arg) {
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    subs r0, r0, #4
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
-; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_neg_offset_bswap:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldr r0, [r0, #-4]
-; CHECK-THUMBv7-NEXT:    rev r0, r0
+; CHECK-THUMBv7-NEXT:    subs r0, r0, #4
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv7-NEXT:    bx lr
 
 
@@ -632,27 +619,14 @@ define i32 @load_i32_by_bswap_i16(i32* %arg) {
 ;
 ; CHECK-THUMBv6-LABEL: load_i32_by_bswap_i16:
 ; CHECK-THUMBv6:       @ %bb.0:
-; CHECK-THUMBv6-NEXT:    ldrb r1, [r0, #2]
-; CHECK-THUMBv6-NEXT:    ldrb r2, [r0, #3]
-; CHECK-THUMBv6-NEXT:    lsls r2, r2, #8
-; CHECK-THUMBv6-NEXT:    adds r1, r2, r1
-; CHECK-THUMBv6-NEXT:    rev r1, r1
-; CHECK-THUMBv6-NEXT:    lsrs r1, r1, #16
-; CHECK-THUMBv6-NEXT:    lsls r1, r1, #16
-; CHECK-THUMBv6-NEXT:    ldrh r0, [r0]
-; CHECK-THUMBv6-NEXT:    rev16 r0, r0
-; CHECK-THUMBv6-NEXT:    adds r0, r1, r0
+; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_bswap_i16:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrh r1, [r0, #2]
-; CHECK-THUMBv7-NEXT:    movw r2, #65535
-; CHECK-THUMBv7-NEXT:    ldrh r0, [r0]
-; CHECK-THUMBv7-NEXT:    rev r1, r1
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv7-NEXT:    rev r0, r0
-; CHECK-THUMBv7-NEXT:    bics r1, r2
-; CHECK-THUMBv7-NEXT:    orr.w r0, r1, r0, lsr #16
 ; CHECK-THUMBv7-NEXT:    bx lr
 
 
@@ -684,20 +658,12 @@ define i32 @load_i32_by_sext_i16(i32* %arg) {
 ;
 ; CHECK-THUMBv6-LABEL: load_i32_by_sext_i16:
 ; CHECK-THUMBv6:       @ %bb.0:
-; CHECK-THUMBv6-NEXT:    ldrb r1, [r0, #2]
-; CHECK-THUMBv6-NEXT:    ldrb r2, [r0, #3]
-; CHECK-THUMBv6-NEXT:    lsls r2, r2, #8
-; CHECK-THUMBv6-NEXT:    adds r1, r2, r1
-; CHECK-THUMBv6-NEXT:    ldrh r0, [r0]
-; CHECK-THUMBv6-NEXT:    lsls r0, r0, #16
-; CHECK-THUMBv6-NEXT:    adds r0, r0, r1
+; CHECK-THUMBv6-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_sext_i16:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrh r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrh r0, [r0, #2]
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #16
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0]
 ; CHECK-THUMBv7-NEXT:    bx lr
   %tmp = bitcast i32* %arg to i16*
   %tmp1 = load i16, i16* %tmp, align 4
@@ -738,12 +704,14 @@ define i32 @load_i32_by_i8_base_offset_index(i8* %arg, i32 %i) {
 ; CHECK-THUMBv6:       @ %bb.0:
 ; CHECK-THUMBv6-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0, #12]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_base_offset_index:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    add r0, r1
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    ldr r0, [r0, #12]
+; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
   %tmp = add nuw nsw i32 %i, 3
   %tmp2 = add nuw nsw i32 %i, 2
@@ -803,12 +771,15 @@ define i32 @load_i32_by_i8_base_offset_index_2(i8* %arg, i32 %i) {
 ; CHECK-THUMBv6-NEXT:    adds r0, r1, r0
 ; CHECK-THUMBv6-NEXT:    movs r1, #13
 ; CHECK-THUMBv6-NEXT:    ldr r0, [r0, r1]
+; CHECK-THUMBv6-NEXT:    rev r0, r0
 ; CHECK-THUMBv6-NEXT:    bx lr
 ;
 ; CHECK-THUMBv7-LABEL: load_i32_by_i8_base_offset_index_2:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    add r0, r1
-; CHECK-THUMBv7-NEXT:    ldr.w r0, [r0, #13]
+; CHECK-THUMBv7-NEXT:    adds r0, r1, r0
+; CHECK-THUMBv7-NEXT:    movs r1, #13
+; CHECK-THUMBv7-NEXT:    ldr r0, [r0, r1]
+; CHECK-THUMBv7-NEXT:    rev r0, r0
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = add nuw nsw i32 %i, 4
@@ -870,7 +841,8 @@ define i32 @zext_load_i32_by_i8(i32* %arg) {
 ; CHECK-THUMBv7:       @ %bb.0:
 ; CHECK-THUMBv7-NEXT:    ldrb r1, [r0]
 ; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #1]
-; CHECK-THUMBv7-NEXT:    orr.w r0, r1, r0, lsl #8
+; CHECK-THUMBv7-NEXT:    lsls r0, r0, #8
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -916,9 +888,10 @@ define i32 @zext_load_i32_by_i8_shl_8(i32* %arg) {
 ; CHECK-THUMBv7-LABEL: zext_load_i32_by_i8_shl_8:
 ; CHECK-THUMBv7:       @ %bb.0:
 ; CHECK-THUMBv7-NEXT:    ldrb r1, [r0]
+; CHECK-THUMBv7-NEXT:    lsls r1, r1, #8
 ; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #1]
 ; CHECK-THUMBv7-NEXT:    lsls r0, r0, #16
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #8
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -965,9 +938,10 @@ define i32 @zext_load_i32_by_i8_shl_16(i32* %arg) {
 ; CHECK-THUMBv7-LABEL: zext_load_i32_by_i8_shl_16:
 ; CHECK-THUMBv7:       @ %bb.0:
 ; CHECK-THUMBv7-NEXT:    ldrb r1, [r0]
+; CHECK-THUMBv7-NEXT:    lsls r1, r1, #16
 ; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #1]
 ; CHECK-THUMBv7-NEXT:    lsls r0, r0, #24
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #16
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -1010,9 +984,10 @@ define i32 @zext_load_i32_by_i8_bswap(i32* %arg) {
 ;
 ; CHECK-THUMBv7-LABEL: zext_load_i32_by_i8_bswap:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrb r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #1]
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #8
+; CHECK-THUMBv7-NEXT:    ldrb r1, [r0, #1]
+; CHECK-THUMBv7-NEXT:    ldrb r0, [r0]
+; CHECK-THUMBv7-NEXT:    lsls r0, r0, #8
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -1057,10 +1032,11 @@ define i32 @zext_load_i32_by_i8_bswap_shl_8(i32* %arg) {
 ;
 ; CHECK-THUMBv7-LABEL: zext_load_i32_by_i8_bswap_shl_8:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrb r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #1]
-; CHECK-THUMBv7-NEXT:    lsls r1, r1, #16
-; CHECK-THUMBv7-NEXT:    orr.w r0, r1, r0, lsl #8
+; CHECK-THUMBv7-NEXT:    ldrb r1, [r0, #1]
+; CHECK-THUMBv7-NEXT:    lsls r1, r1, #8
+; CHECK-THUMBv7-NEXT:    ldrb r0, [r0]
+; CHECK-THUMBv7-NEXT:    lsls r0, r0, #16
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -1106,10 +1082,11 @@ define i32 @zext_load_i32_by_i8_bswap_shl_16(i32* %arg) {
 ;
 ; CHECK-THUMBv7-LABEL: zext_load_i32_by_i8_bswap_shl_16:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrb r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #1]
-; CHECK-THUMBv7-NEXT:    lsls r1, r1, #24
-; CHECK-THUMBv7-NEXT:    orr.w r0, r1, r0, lsl #16
+; CHECK-THUMBv7-NEXT:    ldrb r1, [r0, #1]
+; CHECK-THUMBv7-NEXT:    lsls r1, r1, #16
+; CHECK-THUMBv7-NEXT:    ldrb r0, [r0]
+; CHECK-THUMBv7-NEXT:    lsls r0, r0, #24
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %tmp = bitcast i32* %arg to i8*
@@ -1156,9 +1133,10 @@ define i16 @load_i16_from_nonzero_offset(i8* %p) {
 ;
 ; CHECK-THUMBv7-LABEL: load_i16_from_nonzero_offset:
 ; CHECK-THUMBv7:       @ %bb.0:
-; CHECK-THUMBv7-NEXT:    ldrh r1, [r0]
-; CHECK-THUMBv7-NEXT:    ldrb r0, [r0, #2]
-; CHECK-THUMBv7-NEXT:    orr.w r0, r0, r1, lsl #8
+; CHECK-THUMBv7-NEXT:    ldrb r1, [r0, #2]
+; CHECK-THUMBv7-NEXT:    ldrh r0, [r0]
+; CHECK-THUMBv7-NEXT:    lsls r0, r0, #8
+; CHECK-THUMBv7-NEXT:    adds r0, r0, r1
 ; CHECK-THUMBv7-NEXT:    bx lr
 
   %p1.i16 = bitcast i8* %p to i16*